Pulse - Pain Points

Technical Challenges

LLM Hallucination on Table/Column Names

• Without schema-first enforcement, LLMs invent plausible-sounding table names that don't exist, requiring back-and-forth correction
• Source: Palm Internal (2026-03-09) — Art described this as the #1 problem during the hackathon

Prompt-Level Security is Insufficient

• System prompts can be bypassed by determined users — customer isolation, query restrictions, and data access must be enforced at infrastructure level, not in prompts
• Source: Palm Internal (2026-03-09) — Art: "LLMs can very easily bypass system prompts"

Context Window Pollution

• Mixing data from different customers in the same conversation degrades response quality — context window gets polluted with irrelevant data
• Source: Palm Internal (2026-03-09) — Rodel described this as reason for removing "view as customer" toggle

No Response Evaluation Framework Yet

• Currently no systematic way to evaluate correctness, detect hallucinations, or measure response quality post-chat
• Source: Palm Internal (2026-03-09) — Art: "We do need to check the accuracy, evaluate responses, check for hallucinations at least post-chat"

Cost Scaling Unknown

• Palm Chat uses Palm's own tokens (unlike MCP where user's LLM provider pays). As usage grows, costs will ramp up — need efficiency strategies and LLM evaluation
• Source: Palm Internal (2026-03-09) — Art flagged cost and efficiency as future concern

Product Challenges

No Observability for MCP Usage

• Once MCP access is given, Palm has no visibility into what users query, can't validate responses or improve the feature
• Source: Palm Internal (2026-03-09) — Art: "We lack control once the MCP access is given"

Entity-Level ACLs Create Service Account Complexity

• Current security model uses per-customer service accounts. Adding entity-level access control would require per-customer-per-entity combinations — significant scaling challenge
• Source: Palm Internal (2026-03-09) — Rodel acknowledged this as a tricky scaling problem

Accuracy Disclaimer Needed

• Internal users understand LLM limitations, but external customers need clear messaging that AI responses may contain errors — "Beta" labeling discussed but not yet implemented
• Source: Palm Internal (2026-03-09) — Team discussed whether to show "AI can make mistakes" disclaimer

Adoption Challenges

Users See AI as "Just Another Chatbot"

• Users don't understand AI can be a proactive assistant that surfaces insights they didn't know to ask about — they expect to type a question and get an answer
• Source: Palm Internal (2026-03-10) — Giannis: Amanda didn't know to ask for the categorization analysis; it only happened because CS initiated it

Reactive Model Limits Value Delivery

• Current chat-first model requires users to know what to ask — but the highest-value insights are the ones users don't know they need
• Source: Palm Internal (2026-03-10) — Emma: "We need to be proactive. We need to show the user something worth seeing."

Dashboard Requests Weren't Impactful

• CS felt weekly meetings focused on incremental dashboard changes weren't creating real value for customers
• Source: Palm Internal (2026-03-10) — Giannis: "I had a very transparent and honest conversation with Jen that I was feeling for the past weeks... that we were just attending to new dashboard requests... I felt for some time that it was not that helpful or impactful."